Invisible text watermarks: how companies meet AI Act obligations

Marking AI-generated content sounds like a matter of honesty toward the reader. In practice, it is a technical problem that the industry handles in various ways. heise online describes how Anthropic is extending text watermarking to additional Claude models, set against the backdrop of the EU AI regulation.

According to heise, from 30 September 2026, the invisible watermark will also apply to the output of Claude Fable 5, Claude Sonnet 5, and Claude Opus 4.8. Currently, Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5, and Claude Opus 5 are already watermarked. The last one was marked in a post-launch update. Older models are to be covered by 2 December. This information comes from an email Anthropic sent to customers on 24 September. A key detail: the mark is applied at the model level, and users cannot disable it.

How it works

The method does not rely on hidden characters or metadata, which are easy to strip. Anthropic uses its own variant of the SynthID-Text technique developed by Google DeepMind. The mechanism influences the choice among several equally likely next tokens. Over a longer text, a statistical pattern emerges. With the right key, this pattern can be detected, provided the analyzed passage is long enough. The company claims this does not degrade text quality.

The limitations are more serious than the simplicity of the idea suggests, and Anthropic acknowledges this. A match only indicates that Claude likely participated in producing the text. The model can rewrite an external passage, add its own sentences, or translate someone else's text, and the mark disappears. It is not an evidentiary tool but a probability signal.

Where the obligation comes from

The reason is regulatory. The EU AI Act imposes transparency duties on creators of generative systems. Since 2 August, they must ensure that AI-generated content, including text, is marked in a machine-readable way and recognizable as generated. Anthropic applies its solution globally across all Claude platforms, even though the initiative originated in Europe.

There is also an institutional context. The European Commission runs an AI Office with more than 125 staff across six units. It handles the implementation of the AI Act, including rules for general-purpose models. It can evaluate such models, demand information, and impose sanctions. The marking obligation is just one of many requirements model providers must factor into their production process.

Does it work at all

Criticism comes from two directions. Practitioners point out that text watermarks are vulnerable to ordinary editing and easy to remove via translation or summarization. Conversely, there is the opposite question: what if a mark is detected in a text a person wrote themselves, using a model only for proofreading? Then the meaning of "marked as AI" blurs, making false accusations easy. The safest way to view this mechanism is as a tool for statistical monitoring of content at internet scale, not for determining the authorship of a single paragraph.